Sequential nonlinear encoding: A low dimensional regression algorithm with application to EEG-based driving fatigue detection

Abstract

Regression analysis of real-world data has not always been an easy task, especially when input vectors are presented in a very low dimensional space. EEG-based fatigue detection deals with low dimensional problems and plays a major role in reducing the risk of fatal accidents. We propose a kernel projection pursuit regression algorithm which is a two-step nonlinearity encoding algorithm tailored for such low dimensional problems such as fatigue detection. In this way, data nonlinearity can be investigated from two different perspectives: by first transforming the data into a high dimensional intermediate space and then, applying their spline estimations to the output variables allowing for hierarchical unfolding of data. Experimental results of the SEED VIS database illustrate the average RMSE values of 0.1080% and 0.1054%, respectively, for the temporal and posterior areas of the brain. Our method is validated by conducting some experiments on Parkinson's disease prediction, which further demonstrates the efficiency of our method. This paper proposes a novel regression algorithm to address the encoding problem of highly complex low dimensional data, which is usually encountered in bio-neurological prediction tasks like EEG-based driving fatigue detection.